Method library WORKFLOW / 018

Research workflow/5 steps

Double machine learning for the 401(k) effect (DoubleML)

@doubleml ✓ P D · Jun 29, 2026 · 1.1k views

Source DoubleML — Bach, Chernozhukov, Kurz & Spindler

@misc{doubleml,
  title        = {DoubleML},
  author       = {Bach and Chernozhukov and Kurz and Spindler},
  howpublished = {\url{https://docs.doubleml.org/}},
  note         = {Software / documentation}
}

In the margins · Peer notes@hte_hannah · Pairs beautifully with GRF: DoubleML for the average, causal forest for the heterogeneity. 3

The method workspace

Follow the analysis.

Select a node to explore the method

Method mapInput → 5 steps → output

START WITHInput data

ARRIVE ATResults

Read step 01

Inside the methodReasoning, equations & implementation

Data prep

Build DoubleMLData (y, d, X, z)

Step notesField note & discussion1

Data preparation — shapes the raw inputs into what the estimator expects.

What happens here

Declare outcome, treatment, covariates and (optionally) an instrument.

Reads fromNot specified Feeds into#2

Key code

dml_data = dml.DoubleMLData(df, y_col='net_tfa', d_cols='e401', x_cols=X)

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Data prep

Choose ML learners for the nuisances

Step notesField note & discussion1

Data preparation — shapes the raw inputs into what the estimator expects.

What happens here

e.g. random forest / lasso for E[Y|X] and E[D|X]; estimated by cross-fitting.

Reads from#1 Feeds into#3#4

Key code

ml_l = RandomForestRegressor(); ml_m = RandomForestClassifier()

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Estimation

[DoubleML] Double/debiased ML — PLR / IRM / IIVM

Step notesField note & discussion1

The core estimate — where the causal quantity itself is computed.

What happens here

Partially-linear regression (PLR) for the average effect, with orthogonal scores.

Formula

Y=\theta D+g(X)+\varepsilon,\ \ D=m(X)+v;\quad \psi=(Y-\theta D-g(X))(D-m(X))

The estimator

Double/debiased ML — PLR / IRM / IIVM — Neyman-orthogonal, cross-fitted estimation of treatment effects with arbitrary ML nuisance learners.

ATE Code @ HEAD ↗ Paper ↗

Reads from#2 Feeds into#5

Key code

dml.DoubleMLPLR(dml_data, ml_l, ml_m).fit().summary

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Robustness check

IRM / IIVM cross-checks

Step notesField note & discussion1

A robustness check — does the headline result survive a different lens?

What happens here

Interactive (IRM) and IV (IIVM) models as alternative identifications.

Reads from#2 Feeds into#5

Key code

dml.DoubleMLIRM(dml_data, ml_g, ml_m).fit()

Reference / docs ↗

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Reporting

Coefficient comparison plot

Step notesField note & discussion1

Reporting — turn the numbers into a figure or table a reader can act on.

What happens here

Compare estimates across PLR/IRM/IIVM × learners with 95% confidence intervals.

Reads from#3#4 Feeds intoNot specified

Key code

# error-bar plot of coef ± 1.96·se per model × learner

Discussion on this step (0)

StatsOtterDiscussion starter · Editorial note

Before using this step, make its assumptions and its output explicit. A useful peer note explains what you would check, what could go wrong, or what would change your interpretation.

Questions to explore

Which assumption matters most here, and what evidence could challenge it?
What would you check before passing this output to the next step?

Peer discussion

Bring your experience, a question, or a different perspective.

Data contract

Inputs & data

Outcome (net financial assets), treatment (401(k) eligibility), covariates, and an instrument (eligibility) for the IV model.

Show data format & exampleHide example

Format — one row per household: outcome y, treatment d, covariates X (and instrument z for IIVM).

  net_tfa(y)  e401(d)  age  inc    educ  ...
   12000        1      41  38k    13
    -400        0      53  21k    11

Read the evidence

Results & interpretation

4 figures Workflow discussion 3

01FigureEffect estimates across PLR / IRM / IIVM models and ML learners, with 95% confidence intervals.

02FigureDistribution of net financial assets — the outcome variable.

03FigureTreatment-probability (propensity) distribution by 401(k) eligibility.

04FigureAn overview of the SIPP 401(k) sample used in the example.

Figures reproduced from DoubleML — Bach, Chernozhukov, Kurz & Spindler — unofficial community showcase; all credit to the original authors.

Context & interpretation

The DoubleML 401(k) example (SIPP data). Build the DoubleMLData, choose ML learners, and estimate the effect three ways (PLR / IRM / IIVM). Unofficial summary.

Peer notes

Discuss the workflow (3)

6

@nuisance_nina · Jun 29, 2026

Cross-fitting + Neyman-orthogonal scores = you can throw any ML at the nuisances and still get √n-consistent, valid CIs. This is the way.

5

@aipw_amir · Jun 29, 2026

And it generalizes AIPW cleanly. PLR for a quick read, IRM when effects are nonlinear in X.
2

@kfold_kenji · Jun 29, 2026

The 401(k) example is the perfect teaching case. Same data, three orthogonal scores, sane answers.
2

@hte_hannah · Jun 29, 2026

Pairs beautifully with GRF: DoubleML for the average, causal forest for the heterogeneity.